URL for Download: https://doi.org/10.22032/dbt.52114
This thesis addresses two topics that play a significant role in modern control theory: design of experiments (DoE) and parameter estimation methods for continuous-time (CT) models. In this context, DoE focuses on the impact of experimental design regarding the accuracy of a subsequent estimation of unknown model parameters and applying the theory to real-world applications and its detailed analysis. We introduce the Fisher-information matrix (FIM), consisting of the parameter sensitivities and the resulting highly nonlinear optimization task. By a first-order system, we demonstrate the computation of the information content, its visualization, and an illustration of the effects of higher Fisher information on parameter estimation quality. After that, the topic optimal input design (OID), a subarea of DoE, will be thoroughly explored on the practice-relevant linear and nonlinear model of a 1D-position servo system. Comparison with standard excitation signals shows that the OID signals generally provide higher information content and lead to more accurate parameter estimates using leastsquares methods. Besides, this approach allows taking into account constraints on input, output, and state variables. In the second major topic of this thesis, we treat parameter estimation methods for CT systems, which provide several advantages to identify discrete time (DT) systems, e.g., allows physical insight into model parameters. We focus on modulating function method (MFM) or Poisson moment functionals (PMF) and least squares to estimate unknown model parameters. In the case of noisy measurement data, the problem of biased parameter estimation arises immediately. That is why we discuss the computation and compensation of the so-called estimation bias in detail. Besides the detailed elaboration of a bias compensating estimation method, this work’s main contribution is, based on PMF and least squares for linear systems, the extension to at least slightly nonlinear systems. The derived bias-compensated ordinary least-squares (BCOLS) approach for obtaining asymptotically unbiased parameter estimates is tested on a nonlinear 1D-servo model in the simulation and measurement. A comparison with other methods for bias compensation or avoidance, e.g., total least squares (TLS), is performed. Additionally, the BC-OLS method is applied to the more general MFM. Furthermore, a practical issue of parameter estimation is discussed, which occurs when the system behavior leaves and re-enters the space covered by the identification equation. Using the 1D-servo system, one can show that disabling and re-enabling the PMF filters with appropriate initialization can solve this problem.